OpenAI is swapping out the default model behind ChatGPT, and the headline pitch is simple: it lies less. According to The Verge AI, the company’s new GPT-5.5 Instant model delivers what OpenAI calls “significant improvements in factuality across the board” compared to the GPT-5.3 Instant model it replaces. Rollout starts Tuesday for all ChatGPT users, with GPT-5.3 Instant sticking around as an option for three months before retirement.
The numbers OpenAI is putting on the table are aggressive. Based on internal evaluations cited by The Verge AI, GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law, and finance. On conversations users had flagged for factual errors, inaccurate claims dropped by 37.3%. OpenAI published a system card with more detail on how the evaluations were run.
What’s Actually New
The Verge AI reports GPT-5.5 Instant ships with a stack of upgrades that go beyond the hallucination story:
- Tighter answers. Responses are described as “tighter and more to-the-point,” and the model is told to skip “gratuitous emojis.” That’s a direct response to the chattier, emoji-heavy tone earlier ChatGPT models leaned into.
- Better everyday handling. OpenAI claims the model is “more capable” at routine tasks like analyzing image uploads and judging when to pull a fresh answer from the web instead of guessing.
- Smarter personalization. ChatGPT is now described as “more effective” at pulling context from previous chats and Gmail to shape replies. Google is pushing the same direction with Gemini, so this is shaping up as a core battleground.
- Memory sources. A new feature across all ChatGPT models will surface which pieces of context the chatbot used to personalize a response. Users can delete or correct that information directly.
- Safer model retirement. GPT-5.3 Instant stays selectable for three months. The Verge AI notes users have mourned older model retirements before, so this is a deliberate softer transition.
Who Gets It and When
The base model rolls out to all ChatGPT users starting Tuesday. The enhanced personalization is staged: Plus and Pro users on the web get it first, with mobile apps “coming soon.” Free, Go, Business, and Enterprise tiers are on the roadmap but without a firm date. The memory sources feature is going live now for consumer plans on the web, with mobile to follow.
Why It Matters
Hallucinations are still the single biggest credibility tax on consumer AI. Every confidently wrong answer about a drug interaction, a legal clause, or a financial figure makes enterprise buyers nervous and pushes regulated industries toward narrower tools. A 52.5% drop on high-stakes prompts, if it holds up outside OpenAI’s internal benchmarks, is the kind of number that moves the needle on real adoption in medicine, law, and finance.
What stands out is the framing. OpenAI isn’t pitching GPT-5.5 Instant as smarter. It’s pitching it as more reliable, more concise, and more grounded in your actual context. That’s a maturity move. The frontier race over raw capability is shifting toward a quieter race over trust, persistence, and how well the model knows you across sessions.
The caveat: these are internal evaluations. Independent benchmarks and real-world stress tests will tell the full story over the coming weeks. The system card OpenAI published gives outside researchers something to chew on, but the proof shows up in user-reported errors once it’s in millions of hands.
For anyone running ChatGPT in a workflow that touches sensitive information, the memory sources feature is worth a closer look. Being able to see and edit what the model remembers about you is a meaningful step toward treating AI assistants as accountable tools rather than black boxes. Full details at the original source.